§ Trackr.Live

Qubit Architectures

A qubit is a quantum two-level system. That definition is satisfied by an enormous range of physical systems — superconducting circuits, individual ions, photons, neutral atoms held in optical tweezers, electron spins in semiconductors, exotic quasiparticles, nuclear spins, and several others — and over the last three decades, the quantum computing community has explored most of them in serious engineering programs. No single architecture has yet won; multiple credible candidates remain in active development, with billions of dollars of investment flowing into each of the leading approaches.

This page is the deep-dive companion to the Quantum Computing umbrella page. The scope here is the architectures themselves — the physics that makes each one a candidate, the engineering tradeoffs that distinguish them, the major players pursuing each approach, and the realistic scaling trajectories. Quantum error correction is covered in a separate subpage because it sits on top of any architecture; the algorithmic side is covered in Quantum Algorithms.

What any qubit architecture needs to deliver

Before getting into the specific architectures, it is worth being explicit about what makes a candidate viable. David DiVincenzo’s criteria for a useful quantum computer, formulated in 2000 and refined since, give the canonical framework:

A scalable physical system with well-characterized qubits. The architecture must support adding more qubits without each new qubit fundamentally compromising the ones before it.
The ability to initialize the qubit state. You must be able to reliably put each qubit into a known starting state (typically |0⟩).
Long coherence times relative to gate operation times. The quantum state must survive long enough to execute meaningful algorithms — concretely, the ratio of coherence time to gate time must be large enough that thousands or millions of gate operations can run before the state decoheres.
A universal set of quantum gates. You must have access to enough distinct gate operations to construct arbitrary quantum circuits. The standard universal sets include some single-qubit rotations plus an entangling two-qubit gate.
Qubit-specific measurement. You must be able to measure individual qubits without disturbing the others.

For systems that communicate qubits across distances (relevant to quantum networking and distributed quantum computing), two additional criteria apply:

The ability to interconvert stationary and flying qubits.
The ability to faithfully transmit flying qubits between locations.

Every architecture covered below is evaluated implicitly against these criteria. The architectures differ primarily in how they satisfy each criterion and in which criteria they satisfy easily versus painfully.

Superconducting qubits

Superconducting qubits are the dominant architecture in 2026 by both qubit count and total investment. The approach is pursued by IBM, Google, Rigetti, IQM, the Chinese state research programs, and most of the major academic quantum computing groups.

The physics

A superconducting qubit is a tiny electrical circuit fabricated on a silicon or sapphire substrate. The circuit contains one or more Josephson junctions — thin insulating barriers between two superconducting metals that exhibit quantum-mechanical behavior at low temperatures. The energy levels of the resulting nonlinear oscillator can be engineered such that the two lowest energy states function as the |0⟩ and |1⟩ of a qubit, with the higher excited states sufficiently far away that they can be safely ignored during operation.

The dominant qubit design in production systems is the transmon, introduced in 2007 by Robert Schoelkopf’s group at Yale. The transmon trades reduced charge-noise sensitivity for moderate anharmonicity (the energy gap between |0⟩→|1⟩ and |1⟩→|2⟩) and has been the standard design across the industry for the last decade. Variants — fluxonium, the Andreev qubit, and several others — are pursued in research settings but have not yet displaced the transmon in production systems.

Qubits are coupled together through capacitive or inductive elements, with the coupling strength engineered to enable specific two-qubit gates. The most common entangling gates in superconducting systems are the cross-resonance gate (used by IBM) and the iSWAP family of gates (used by Google), each producing entanglement through different microwave drive protocols.

The operating environment

Superconducting qubits operate at roughly 15 millikelvin — colder than deep space, achievable only with dilution refrigerators that use a mixture of helium-3 and helium-4 isotopes. The fridge is the most visually distinctive piece of any superconducting quantum computer: a multi-stage gold-plated structure suspended in a vacuum chamber, with each stage at progressively colder temperatures down to the millikelvin “mixing chamber” where the qubits sit.

The microwave control signals that drive gate operations come from room-temperature electronics and are routed down through the fridge stages on coaxial cables that get progressively colder and more attenuated. Each qubit needs at least one control line and one readout line; large systems require tens of thousands of these lines, and managing the wiring fanout (without overwhelming the fridge’s cooling capacity) is one of the dominant engineering challenges.

Performance numbers

Typical 2026 transmon performance:

Gate times: 20-200 nanoseconds for single-qubit gates, 50-500 ns for two-qubit gates.
Coherence times: T1 (relaxation) of 100-500 microseconds, T2 (dephasing) of 50-300 microseconds. The best published numbers in research devices exceed 1 millisecond; production-scale systems run lower.
Gate fidelities: 99.9%+ for single-qubit gates, 99.0-99.7% for two-qubit gates in leading systems. The error correction threshold for the surface code is roughly 99% for two-qubit gates, so the leading systems are above the threshold for fault-tolerant operation.

The ratio of coherence time to gate time gives roughly 1,000-10,000 gate operations per coherence time. This is enough for current NISQ algorithms but well below what fault-tolerant cryptographically-relevant algorithms require — the gap is bridged by quantum error correction rather than by raw coherence improvements.

The major players and notable systems

IBM has pursued superconducting qubits since the early 2010s and operates the largest publicly accessible quantum computers through their IBM Quantum cloud. Notable systems:

Eagle (127 qubits, 2021) — the first IBM system above 100 qubits.
Osprey (433 qubits, 2022) — pushed qubit count further but with mixed quality.
Condor (1,121 qubits, 2023) — the largest single-chip superconducting processor at announcement.
Heron (133 qubits initially, expanded to 156 in Heron r2 in 2024) — focused on higher fidelity rather than higher count. Heron processors are the basis for IBM’s current production quantum systems.
IBM Quantum System Two (2024) — modular architecture connecting multiple Heron processors with cryogenic couplers, designed for the longer-term scaling roadmap.

IBM’s announced roadmap continues with Flamingo (1,386 qubits across multiple chips, planned 2025-2026), Kookaburra (4,158 qubits across three Flamingo processors), and Blue Jay (~10,000 logical qubits, targeted 2033). The roadmap explicitly transitions from raw physical qubit counts to logical qubit counts as error correction matures.

Google has pursued superconducting qubits since acquiring John Martinis’s group from UC Santa Barbara in 2014. Notable systems:

Sycamore (53 qubits, 2019) — the system used for the “quantum supremacy” demonstration in late 2019, a random circuit sampling task that Google claimed would take a classical computer 10,000 years. Subsequent classical algorithm improvements reduced the classical cost substantially, but the demonstration remains a notable milestone.
Willow (105 qubits, December 2024) — Google’s announcement that Willow demonstrated below-threshold quantum error correction for the first time, where increasing the surface code distance reduced the logical error rate rather than increasing it. This is the threshold that quantum error correction theory has predicted for thirty years, finally crossed in hardware. The demonstration was at a single logical qubit at distance 3, 5, and 7; scaling to many logical qubits remains future work.

Google’s research focus has been less on maximizing qubit count and more on per-qubit quality and error correction demonstrations. The Willow result is widely considered the most significant near-term-future-relevant quantum computing milestone of 2024.

Rigetti Computing, IQM (Finland), OQC (UK), Anyon Systems (Canada), and several others operate smaller superconducting systems. Origin Quantum (China) and other Chinese efforts pursue similar architectures with limited international visibility.

Scaling challenges

The dominant scaling challenges for superconducting architectures:

Wiring fanout. Every additional qubit requires additional control and readout lines, and the fridge’s cooling capacity limits how many lines can be brought in. Multiplexed control schemes and on-chip electronics (cryo-CMOS) are active research areas.
Cross-talk and frequency crowding. Transmons are typically operated at fixed frequencies, and as more qubits are added to a chip, frequency collisions and unintended interactions become harder to avoid.
Connectivity. Superconducting qubits naturally couple only to their physical neighbors on the chip, giving the typical “heavy-hex” or square-lattice connectivity. Algorithms that need all-to-all connectivity require many SWAP operations to move qubits around, adding gate count and accumulated error.
Modular scaling. Single chips have practical fabrication limits; scaling beyond requires reliable inter-chip couplers, which adds error and latency. IBM Quantum System Two and Google’s modular roadmap address this directly.

The trajectory is positive — qubit quality, count, and error correction demonstrations have all advanced substantially in 2023-2025 — but the gap to cryptographically-relevant scale remains large.

Trapped-ion qubits

Trapped-ion systems represent the longest-running architecture pursuit in the field, with origins in atomic physics work from the 1970s and 1980s. The architecture has consistently delivered the highest individual qubit fidelities, the longest coherence times, and the cleanest all-to-all connectivity in any quantum platform, at the cost of slower gate operations and a different set of scaling challenges than superconducting systems.

The physics

A trapped-ion qubit is an individual atomic ion held in vacuum by electric and magnetic fields, with internal energy states used to encode the qubit. The most common ion species in production systems are ytterbium-171 (used by IonQ and Quantinuum), barium-137 (used in some Quantinuum systems), and calcium-43 (used by AQT and various academic groups). The choice of ion affects the operating wavelengths, hyperfine structure, and gate mechanisms but is fundamentally a matter of engineering optimization rather than a physics-level distinction.

Gate operations are performed with lasers: individual ion addressing is done with focused laser beams that drive specific transitions between energy levels. Two-qubit gates rely on the shared motional modes of the ion crystal — driving the ions in ways that couple their spin states through their shared vibrational motion. The most common two-qubit gate is the Mølmer-Sørensen gate, which produces entanglement through a state-dependent force.

The defining advantage of trapped-ion qubits is that all qubits in a single trap can directly interact with all other qubits through the shared motional modes. There is no nearest-neighbor restriction. An algorithm that needs to entangle qubit 1 with qubit 50 can do so directly without intermediate SWAP operations.

The operating environment

Trapped-ion systems do not require dilution refrigerators. The ions are cooled to microkelvin temperatures using laser cooling techniques, but the apparatus around them operates near room temperature, with the ions sitting in ultra-high vacuum chambers (typically 10⁻¹¹ Torr or better) to prevent collisions with background gas.

The technical infrastructure is dominated by lasers — typically multiple stabilized lasers per ion species, locked to extremely precise reference frequencies. The laser systems are physically larger than the trap itself and represent the bulk of the engineering complexity.

Performance numbers

Typical 2026 trapped-ion performance:

Gate times: 10-100 microseconds for single-qubit gates, 100-1000 microseconds for two-qubit gates. Roughly 1,000x slower than superconducting gates.
Coherence times: seconds to minutes for hyperfine qubits, with some demonstrations extending to hours under ideal conditions. Dramatically longer than superconducting coherence.
Gate fidelities: 99.99%+ for single-qubit gates, 99.9%+ for two-qubit gates in leading systems. The highest published two-qubit fidelities in any quantum platform have come from trapped-ion experiments.

The gate-time-to-coherence-time ratio is roughly 100,000-1,000,000, comparable to or better than superconducting despite the slower gates. The intrinsic quality is high; the engineering challenge is scaling.

The major players

IonQ is the leading commercial trapped-ion company, with publicly traded stock and systems available through major cloud providers. Notable systems:

Forte (32 qubits) — the production system through 2023-2024.
Forte Enterprise — productized variant.
Tempo (announced, targeting 64+ algorithmic qubits) — next-generation system in development.

IonQ uses the algorithmic qubits (AQ) metric as their primary capability measure, which attempts to capture the number of qubits that can be used together effectively in an algorithm rather than the raw count. AQ counts have grown from 11 to 35 between 2020 and 2025 across IonQ’s systems.

Quantinuum (the merger of Honeywell Quantum Solutions and Cambridge Quantum) operates the highest-fidelity trapped-ion systems in the world as of 2026. Their H2 (56 qubits) system has demonstrated multiple record-setting results, including the first demonstration of fault-tolerant logical qubit operations exceeding the fidelity of the underlying physical qubits — a critical milestone for the path to fault-tolerant quantum computing.

AQT (Alpine Quantum Technologies, Austria) operates trapped-ion systems with calcium-43 ions, primarily for research deployment.

Universal Quantum (UK) is pursuing a modular trapped-ion architecture aimed at scalability through interconnected microtraps.

Scaling challenges

The dominant scaling challenges for trapped-ion architectures are different from superconducting:

Per-trap qubit limit. Loading more ions into a single linear trap eventually causes the motional modes to become spectrally crowded, making high-fidelity gates difficult. Current single-trap systems operate with tens of ions; scaling beyond requires moving ions between trap zones or using fundamentally different trap geometries.
QCCD (Quantum Charge-Coupled Device) architectures. Quantinuum’s approach to scaling: physically shuttle ions between specialized zones for gates, storage, and measurement. The architecture is operationally complex but provides a clean scaling path.
Photonic interconnects between traps. IonQ and several research groups are developing photonic networking between separate ion traps, allowing modular scaling beyond a single trap. The interconnect fidelity remains a bottleneck.
Gate speed. Trapped-ion gates are inherently slower than superconducting gates. For algorithms running millions of operations, the total runtime becomes significant.

The trapped-ion community’s path to scale runs through modular architectures rather than through larger single traps. Whether that path can match the qubit counts of superconducting systems on a relevant timeline is a contested question.

Photonic qubits

Photonic quantum computing uses individual photons as qubits, with quantum information encoded in properties like polarization, time-bin, or path. The architecture has theoretical advantages that have made it a persistent candidate for fault-tolerant quantum computing, but the engineering has been substantially harder than the early theoretical work suggested.

The physics

A photonic qubit is typically encoded in the dual-rail representation: a single photon present in one of two optical modes encodes |0⟩ and |1⟩ depending on which mode it’s in. Other encodings — polarization (horizontal vs vertical), time-bin (early vs late), and continuous-variable encodings using squeezed light — are used in various implementations.

Single-qubit gates are straightforward: linear optical components (beam splitters, phase shifters) can perform arbitrary single-qubit operations on photonic qubits. The hard part is two-qubit gates. Photons do not naturally interact with each other in linear media, which means the standard approaches for entangling photonic qubits all rely on indirect mechanisms:

The KLM scheme (Knill-Laflamme-Milburn, 2001) uses measurement-induced nonlinearities — projecting the joint state of multiple photons after they interact at beam splitters — to produce entangling gates probabilistically.
Fusion-based quantum computing (Bartolucci et al., 2021) uses photon fusion measurements to build up large entangled states from small “resource states,” with the architecture designed around the probabilistic nature of photonic gates.
Continuous-variable approaches use squeezed light and homodyne detection to perform gate operations on the quadratures of optical modes, with different security and scaling properties than discrete-variable schemes.

Operating environment

Photonic quantum computers can in principle operate at room temperature, which is genuinely different from every other architecture covered here. The optical components themselves don’t require cryogenic cooling. In practice, single-photon detectors (typically superconducting nanowire single-photon detectors, SNSPDs) do require cryogenic cooling, but the cooling requirements are less extreme than the dilution-fridge regime of superconducting qubits.

The engineering challenge is photon loss. Every optical component (waveguide, coupler, detector) has some probability of losing the photon, and the loss budget compounds across the components in a circuit. Modern integrated photonic platforms have been steadily reducing loss, but the loss tolerance remains a fundamental architectural constraint.

Performance characteristics

Photonic quantum computing performance is harder to summarize than other architectures because the relevant metrics are different. There is no concept of “coherence time” in the same sense — photons either survive or don’t, and a surviving photon doesn’t degrade over time the way a matter-based qubit does. The relevant metrics are photon source brightness, component loss, detector efficiency, and the rate at which resource states can be produced.

For the leading photonic approaches in 2026, the operational picture is:

Resource state generation is the bottleneck. Producing entangled photonic states reliably and at high rate is harder than the early theoretical work suggested.
Two-qubit gates are probabilistic. Even with idealized components, photonic gates succeed only some fraction of the time. The architecture must accommodate this through redundancy and post-selection.
Integration is advancing rapidly. Modern photonic chips integrate many components on silicon photonics or lithium niobate platforms, dramatically reducing loss and improving stability.

The major players

PsiQuantum is the largest and most ambitious photonic quantum computing company, pursuing a fault-tolerant million-qubit architecture from the outset. The company has not publicly demonstrated working quantum computers; their pitch is that the architecture they’re building is the right one for scale, and that intermediate-scale demonstrations are not the goal. The bet is large — billions of dollars of investment — and the timeline is long.

Xanadu (Canada) has built photonic quantum computers using continuous-variable encoding, with cloud access available through their platform. Xanadu published a quantum advantage demonstration in 2022 using their Borealis system on a Gaussian boson sampling task.

ORCA Computing (UK) focuses on photonic quantum computing for near-term commercial applications, particularly in machine learning contexts.

QuiX Quantum (Netherlands) builds photonic processors for boson sampling and related applications.

Scaling considerations

Photonic quantum computing has a different scaling profile than matter-based architectures. The qubit count is less meaningful as a metric because photonic qubits are produced and consumed continuously during computation — there is no fixed array of “the qubits.” The relevant scaling questions are about resource state production rate, photon loss budget, and the size of the entangled states that can be reliably constructed.

If the engineering challenges can be overcome, photonic architectures have a plausible path to very large scale because the underlying optical components are compatible with modern semiconductor fabrication. Whether the engineering will get there on a relevant timeline is genuinely unclear.

Neutral atom qubits

Neutral atom quantum computing has been the fastest-rising architecture in the 2020s, with multiple companies and research groups demonstrating systems that scaled from tens to thousands of qubits in a handful of years. The approach combines some of the best features of trapped-ion systems with a more favorable scaling path.

The physics

A neutral atom qubit is an individual atom — typically rubidium-87 or cesium-133 — held in a focused laser beam called an optical tweezer. The atom is cooled to microkelvin temperatures via laser cooling and then trapped at the tweezer focus. Qubit states are encoded in hyperfine ground-state sublevels of the atom, which have excellent coherence properties.

Single-qubit gates use focused laser pulses that drive specific transitions between hyperfine levels. Two-qubit gates use Rydberg interactions: temporarily exciting one or both atoms to highly excited Rydberg states, where the atoms experience strong long-range interactions that produce entanglement. The Rydberg state has a short lifetime but a strong interaction, allowing fast entangling gates.

The defining advantage of neutral atom systems is reconfigurability: the optical tweezer pattern can be reshaped between operations, physically moving atoms to bring different pairs into Rydberg interaction range. This produces effectively all-to-all connectivity with high flexibility, without requiring the ion-shuttling complexity of QCCD architectures.

Operating environment

Neutral atom systems operate in ultra-high vacuum chambers, similar to trapped-ion systems, but without the requirement for the elaborate ion-trap electrode structures. The apparatus is dominated by lasers — multiple stabilized lasers for cooling, trapping, and qubit control — and by the optical systems that produce the tweezer arrays.

Like trapped-ion systems, the atoms themselves are at microkelvin temperatures via laser cooling, but the surrounding apparatus operates near room temperature.

Performance numbers

Typical 2026 neutral atom performance:

Gate times: 100 nanoseconds to a few microseconds for single-qubit gates, similar range for two-qubit Rydberg gates. Faster than trapped-ion gates, slower than superconducting.
Coherence times: seconds for hyperfine qubits, comparable to trapped-ion.
Gate fidelities: 99.5%+ single-qubit fidelities, 99.0-99.5% two-qubit fidelities in leading systems. Below the best trapped-ion fidelities but improving rapidly.

The combination of fast gates, long coherence, and flexible connectivity has made neutral atom systems competitive with the established architectures despite being a newer entrant.

The major players

Atom Computing announced a 1,180-atom system in October 2023, the first quantum computer of any architecture to exceed 1,000 qubits. The system used a 2D array of neutral atoms with high coherence times and demonstrated coherent operations across the full array. Atom Computing’s roadmap targets scalable error correction on the same architecture.

QuEra Computing operates 256-atom and larger systems through cloud access on AWS Braket and other platforms. QuEra has published notable results on logical qubit operations and quantum advantage demonstrations on combinatorial optimization problems.

Pasqal (France) pursues neutral atom quantum computing with a focus on European quantum infrastructure deployments, with systems at several European supercomputing centers.

Infleqtion (formerly ColdQuanta) builds neutral atom systems and supporting quantum technology infrastructure, including optical clocks and quantum sensors alongside the quantum computing work.

Scaling characteristics

Neutral atom architectures have several scaling advantages that have made the field’s recent progress possible:

Atom arrays scale naturally. The optical tweezer pattern can be made larger by using larger optical systems and more sophisticated holographic generation. Going from 100 to 1,000 atoms is engineering work, not a new architectural concept.
No wiring fanout. Unlike superconducting systems, neutral atom architectures do not require an individual control line per qubit. The same laser system can address many atoms through optical multiplexing.
Connectivity is reconfigurable. Atoms can be moved between operations, providing effectively all-to-all connectivity without intermediate SWAP overhead.
Loss and refilling. Atoms can be lost to background gas collisions, but the system can detect and refill missing atoms continuously during operation.

The neutral atom community has been the most surprising part of the field over the last three years. The architecture went from “interesting research direction” to “competitive with established platforms” faster than most observers predicted, and the trajectory remains favorable.

Topological qubits

Topological quantum computing is the long-running architectural bet that, if it works, would dramatically reduce the error correction overhead by encoding qubits in topologically protected quantum states. The theoretical appeal has been clear since the late 1990s; the engineering reality has been substantially harder than the theory suggested.

The physics

A topological qubit would encode quantum information in the non-local properties of exotic quasiparticles called non-abelian anyons, particularly Majorana zero modes at the ends of certain semiconductor-superconductor heterostructures. The information would be stored in the braiding pattern of the anyons rather than in the local state of any specific particle, which would make it intrinsically resistant to local noise sources.

The promise is that topological qubits would have error rates orders of magnitude lower than other architectures, potentially eliminating the need for the resource-intensive surface code error correction. A topological quantum computer that delivered on its theoretical promise would require dramatically fewer physical qubits than other architectures to reach cryptographically-relevant scale.

The state of the field

Microsoft has pursued topological quantum computing as a strategic bet since the early 2010s, with substantial investment in research at Microsoft Station Q and partner institutions. The track record has been controversial.

A 2018 Nature paper by a Microsoft-affiliated group claimed observation of Majorana zero modes in a semiconductor-superconductor device. The paper was retracted in 2021 after concerns were raised about data interpretation and selective presentation. The retraction was a substantial embarrassment for the topological quantum computing program.

In February 2025, Microsoft announced the Majorana 1 processor, claiming the first topological qubit prototype based on a new “topoconductor” device structure. The announcement was accompanied by claims of stable Majorana mode detection and a roadmap to scale. The claims have been received with skepticism by parts of the physics community, with several researchers publishing technical critiques arguing that the announced measurements do not unambiguously establish the presence of Majorana modes. As of mid-2026, the status remains contested — Microsoft maintains the claims, several independent researchers dispute them, and the broader community is waiting for additional results that would settle the question.

For practical purposes in 2026, topological qubits are not yet an engineering reality. They remain a research bet that may pay off on a longer timeline. The other architectures covered on this page are where current production systems are built.

Spin qubits in silicon

Spin qubits use the spin state of individual electrons (or nuclei) confined in semiconductor quantum dots as qubits. The architecture has a compelling industrial argument: it could potentially leverage existing silicon semiconductor manufacturing infrastructure to produce quantum processors using the same fabs and processes that produce classical chips.

The physics

A spin qubit is an electron (or pair of electrons, or a nuclear spin) confined in a quantum dot — a tiny electrostatic well formed in a semiconductor heterostructure. The qubit state is encoded in the spin direction (up vs down), with single-qubit operations performed via microwave magnetic fields or electric dipole spin resonance.

Two-qubit gates exploit the exchange interaction between neighboring electrons. The interaction is strong, fast, and naturally short-range, providing a path to fast two-qubit gates between physically adjacent qubits.

Silicon is the dominant host material because isotopically purified silicon-28 has zero nuclear spin, dramatically reducing the magnetic noise that limits coherence. Pure silicon-28 substrates have demonstrated coherence times exceeding a second for electron spin qubits — competitive with the best trapped-ion numbers.

The major players

Intel has pursued spin qubits since the late 2010s as part of their Quantum Computing program, with systems demonstrating 12-qubit “Tunnel Falls” processors fabricated in their commercial silicon fabs. The strategic argument is that Intel can leverage decades of silicon manufacturing expertise to scale spin qubit systems faster than competitors building dedicated quantum-specific fabrication infrastructure.

Diraq (Australia) has demonstrated multi-qubit spin qubit systems with high single-qubit fidelities, leveraging research from the University of New South Wales.

Quantum Motion (UK) and Silicon Quantum Computing (Australia) are pursuing similar architectures with different specific approaches to qubit design.

Scaling considerations

The argument for spin qubits is fundamentally an industrial scaling argument: if classical chip manufacturing can fabricate billions of transistors per chip, the same processes (with appropriate adaptations) might be able to fabricate millions of spin qubits per chip. The realization of this argument has been slower than the optimistic projections — spin qubits remain at smaller qubit counts than superconducting or neutral atom systems — but the long-term scaling promise is unique among the architectures.

Spin qubits face their own challenges: single-qubit and two-qubit gate fidelities lag the leading platforms, control electronics for large arrays are not yet demonstrated, and the move from research-grade fabrication to true volume production is not yet evidenced. The architecture remains a credible long-term contender rather than a current-leader.

Other approaches worth knowing

A handful of additional architectures appear in research contexts but have not yet reached the production scale of those covered above:

Nitrogen-vacancy (NV) centers in diamond use defect states in crystalline diamond as qubits. Excellent coherence properties at room temperature, but scaling beyond a few qubits has been challenging. NV centers are more competitive for quantum sensing applications than for quantum computing.

Nuclear spin systems use the spin states of atomic nuclei as qubits, with even longer coherence times than electron spin systems. Liquid-state NMR was one of the earliest demonstrated quantum computing platforms but has fundamental scaling limits. Solid-state nuclear spin systems remain a research direction.

Cat qubits (used by Alice & Bob, AWS Quantum) are bosonic qubits encoded in superpositions of coherent states of a superconducting cavity. The architecture provides intrinsic protection against certain error types and is a hybrid approach combining superconducting circuit technology with bosonic encoding.

Trapped electrons (in Penning traps and similar) are pursued by a small number of groups as a potential alternative to trapped ions with different operational characteristics.

Architecture comparison

A summary table of the leading architectures in 2026:

Architecture	Best gate fidelity	Coherence time	Gate time	Connectivity	Operating temp	Scaling path
Superconducting	99.7% (2Q)	100-500 μs	50-500 ns	Nearest neighbor	15 mK	Modular chips
Trapped ion	99.9%+ (2Q)	Seconds	10-1000 μs	All-to-all	Room temp*	QCCD / photonic interconnect
Photonic	Hard to compare	“Infinite”	~ns	Reconfigurable	Room temp*	Fusion-based architecture
Neutral atom	99.5% (2Q)	Seconds	100-1000 ns	Reconfigurable	Room temp*	Larger atom arrays
Topological	N/A	N/A	N/A	N/A	mK	Unproven
Spin (silicon)	99% (2Q)	~ms-s	10-1000 ns	Nearest neighbor	mK	Silicon fab compatibility

*Atoms/ions/photons themselves are at microkelvin via laser cooling; surrounding apparatus is near room temp.

What might win

The honest answer to “which architecture will reach cryptographically-relevant scale first” is that we don’t know, and the most credible practitioners in the field are unwilling to predict confidently. A few observations that should inform reasonable expectations:

Multiple architectures will likely cross meaningful milestones in parallel. The field is broad enough and well-funded enough that multiple approaches will continue to advance. The relevant question is which architectures cross particular thresholds (1,000 logical qubits, 10,000 logical qubits, 1 million logical qubits) on which timelines, not which architecture is “the winner.”

Superconducting has the most accumulated investment and the largest installed base. This matters for ecosystem effects — the software stacks, the cloud platforms, the algorithmic research is disproportionately developed against superconducting hardware. Other architectures have to overcome ecosystem inertia even if their underlying physics is favorable.

Neutral atoms have the most favorable recent trajectory. The architecture went from research curiosity to thousands-of-qubits in a few years, and the scaling path looks genuinely good. Whether the trajectory continues is a question for the next few years.

Trapped ions have the highest individual qubit quality. For applications where fidelity matters more than count, trapped-ion systems are the current leaders, and the modular scaling path may eventually reach competitive total qubit counts.

Photonic systems have a fundamentally different scaling profile. PsiQuantum’s million-qubit bet either pays off dramatically or doesn’t — there is less middle ground than with other architectures.

Topological qubits remain a wild card. If Microsoft’s claims hold up, the architecture could change the field dramatically. If they don’t, it’s a research direction that has consumed a decade of effort without producing engineering results. The next several years should resolve which case applies.

Architecture-specific error correction implications

The physical qubit count needed to produce a useful logical qubit varies meaningfully across architectures. The factors that drive the difference:

Physical qubit error rate determines the surface code distance required for a given logical error rate. Lower physical error rates require smaller codes, fewer physical qubits per logical qubit.

Connectivity affects the efficiency of the error correction code. Architectures with all-to-all connectivity (trapped ion, reconfigurable neutral atom) can use error correction codes that are more efficient than the standard surface code, which is designed for nearest-neighbor connectivity.

Native gate set affects how many physical gates are required to implement the logical operations. Some architectures have native gates that map naturally to error correction operations; others require additional decomposition.

The implication is that a “1,000 physical qubit” superconducting system, a “1,000 physical qubit” trapped-ion system, and a “1,000 physical qubit” neutral atom system are not directly comparable in terms of useful computational capacity. The comparison that matters is logical qubits at a given logical error rate, and that calculation depends on the full architectural picture.

The Quantum Error Correction subpage covers the codes and the threshold theorem in detail.

State of the art, mid-2026

A snapshot of where the field stands as of mid-2026:

Largest published superconducting system: IBM Heron variants and Condor at the high count; Google Willow with the strongest error correction demonstration.
Largest published trapped-ion system: Quantinuum H2 at 56 qubits with the highest published two-qubit fidelities in any production platform.
Largest published neutral atom system: Atom Computing’s 1,180-atom array (announced 2023); subsequent systems have advanced further.
Photonic milestones: Xanadu’s Borealis demonstration (2022) and continuing platform development; PsiQuantum still pre-production.
Topological qubits: Microsoft’s Majorana 1 announcement (February 2025), contested as noted above.
Quantum advantage demonstrations: Google Sycamore (2019), USTC Jiuzhang and Zuchongzhi (2020-2021), Xanadu Borealis (2022). All on contrived problems with no practical application.
Logical qubit demonstrations: Google Willow’s below-threshold result (December 2024), Quantinuum’s fault-tolerant operations, multiple groups demonstrating distance-3 to distance-7 codes.

The pace of advancement has been substantial across all major architectures. The gap to cryptographically-relevant scale remains large but the trajectory is favorable. The honest summary: nothing in 2026 breaks RSA-2048, the post-quantum cryptography migration is being driven by future-looking risk assessment rather than present capability, and the next decade is going to see substantial further advancement in qubit counts, qubit quality, and error correction.

Where to go next on this site

Adjacent material on this site:

Quantum Computing — the umbrella overview covering the fundamental concepts and the relationship to cryptography.
Quantum Algorithms — the algorithmic side: what Shor’s, Grover’s, and the other quantum algorithms actually do, and which problems they solve.
Quantum Error Correction — surface codes, the threshold theorem, fault tolerance, the bridge from physical to logical qubits.
Quantum Hardware: State of the Art — the current production systems in 2026, vendor by vendor, with capability metrics.
Post-Quantum Cryptography — the cryptographic response to the quantum threat.

The architectures landscape is the part of quantum computing that has been moving fastest in 2023-2026, and the page will need updating as the field advances. The goal is to keep this as a current reference rather than a fixed snapshot — corrections and updates are part of the page’s life.